BDM-YOLOv8n: A high-performance model for accurate fire detection in aerial imagery

In recent years, UAV aerial imagery has emerged as a pivotal tool in fire detection. However, when capturing images at long distances, it will be affected by factors such as limited viewing angles, complex backgrounds and environmental interference, leading to reduced accuracy and reliability in fire detection. To solve the above problems, this paper proposes the BMD-YOLOv8n model. First, a comprehensive dataset, AeroFlame, was constructed. In the model, a Bidirectional Feature Pyramid Network with P2 layer (P2-BiFPN) was designed to fuse features of different scales and better retain the information of small targets. Next, A module integrating Diverse Branch Block, Cross Stage Partial and Efficient Layer Aggregation Network (DBBCSPELAN) was designed to enhance the feature extraction ability. To meet the feature extraction requirements while improving the computational efficiency and reducing background noise, a Simple Attention Module (SimAM) was introduced. Finally, the standard CIoU loss function was replaced by the Multi-Perspective Distance-IoU (MPDIoU) loss function to improve the accuracy of boundary box regression.The experimental results show that BDM-YOLOv8n outperforms baseline YOLOv8n by 3.5 %, 3.7 %, and 3.5 % in precision, recall, and mAP50, respectively. In addition, on two single-scene open-source fire datasets, FLAM and FASDD, the proposed fire detection model's detection accuracy reaches 92.4 % and 91.8 %, respectively.